Abstract: Zero-shot image recognition aims to address recognition issues in the absence of samples. To address the issues of insufficient discriminative features for target samples and domain shift in zero-shot image recognition, a zero-shot image recognition method based on redundancy reduction and dual perception constraints is proposed on the basis of generative models. To tackle the issue of insufficient discriminative features for target samples, the loss constraint of visual center is firstly introduced into the generator to make the generated pseudo-features between classes more compact, thereby improving the inter-class differentiation of pseudo-features. Additionally, a redundancy reduction module is added after the generator to reduce interference from redundant information and emphasize discriminative feature information. To address the domain shift problem, a dual perception regularization constraint is proposed to enforce dual perception constraint on both real visual features and pseudo-visual features, making the generated pseudo-features align with the real ones more closely. Furthermore, a cycle consistency loss is utilized to perform semantic reconstruction on pseudo-visual features generated by the semantic decoder, further alleviating domain shift. The effectiveness of the proposed method was verified by the experiments on the AWA, CUB, SUN and FLO datasets. The method was also applied to the zero-shot image retrieval task, and the experimental results showed that the proposed method had good generalization performance, thus easy to extend to other applications.